Intelligent Detection and Identification in Fiber-Optical ... · Huijuan WU et al.: Intelligent Detection and Identification in Fiber-Optical Perimeter Intrusion Monitoring System

PHOTONIC SENSORS / Vol. 5, No. 4, 2015: 365–375

Intelligent Detection and Identification in Fiber-Optical Perimeter Intrusion Monitoring System Based on the FBG

Sensor Network

Huijuan WU1*, Ya QIAN1, Wei ZHANG1, Hanyu LI2, and Xin XIE1

1 Key Laboratory of Optical Fiber Sensing and Communications, Ministry of Education, University of Electronic Science

and Technology of China, Chengdu, 611731, China 2 Chinese People’s Liberation Army (CPLA) Urumqi Institute of the Army, Urumqi, 830042, China *Corresponding author: Huijuan WU E-mail: [email protected]

Abstract: A real-time intelligent fiber-optic perimeter intrusion detection system (PIDS) based on the fiber Bragg grating (FBG) sensor network is presented in this paper. To distinguish the effects of different intrusion events, a novel real-time behavior impact classification method is proposed based on the essential statistical characteristics of signal’s profile in the time domain. The features are extracted by the principal component analysis (PCA), which are then used to identify the event with a K-nearest neighbor classifier. Simulation and field tests are both carried out to validate its effectiveness. The average identification rate (IR) for five sample signals in the simulation test is as high as 96.67%, and the recognition rate for eight typical signals in the field test can also be achieved up to 96.52%, which includes both the fence-mounted and the ground-buried sensing signals. Besides, critically high detection rate (DR) and low false alarm rate (FAR) can be simultaneously obtained based on the autocorrelation characteristics analysis and a hierarchical detection and identification flow.

Keywords: Behavior impact classification, fiber-optical fence, PIDS, security, FBG

Citation: Huijuan WU, Ya QIAN, Wei ZHANG, Hanyu LI, and Xin XIE, “Intelligent Detection and Identification in Fiber-Optical Perimeter Intrusion Monitoring System Based on the FBG Sensor Network,” Photonic Sensors, 2015, 5(4): 365–375.

1. Introduction

With an increase in terrorism in recent years

especially after 911, it brings most important yet difficult security challenges globally, such as the perimeter protection of airports, railway stations,

government buildings, and military bases. The advent of fiber sensors opens up more opportunities to perimeter security and provides a new promising

solution for this application [1–4]. Comparing with those conventional perimeter intrusion detection systems (PIDSs) that use ultrasonic, radar,

microwave, and infrared or photo-electric sensors to detect intrusions, the optical fiber sensor (OFS) has outstanding advantages of passive operation, high

sensitivity, good reliability in harsh conditions, long-distance capability, electro-magnetic interference immunity (EMI), and corrosion

resistance, etc. In particular, OFS does not need any power supply along the fiber link, and hence it is an ideal choice for long or medium long distance

applications in harsh field environments. Typical OFS technologies in the security area include the phase-sensitive optical time domain reflectometry

Received: 22 August 2015 / Revised: 15 September 2015 © The Author(s) 2015. This article is published with open access at Springerlink.com DOI: 10.1007/s13320-015-0274-8 Article type: Regular

Photonic Sensors

366

(Φ-OTDR) [5–7] and many other OTDR technologies [8], Sagnac, Michelson, Mach-Zehnder (M-Z) and their combination structures [9–12], white light interferometers[13], and

quasi-distributed FBG sensor networks [14–19].

In comparison with the highly sensitive OTDR

and the interferometer-based fiber fences, fiber

Bragg grating (FBG) sensors are immune to the

environmental interferences and thus with more

reliable detection accuracy and much lower nuisance

alarm rates (NARs). Besides, they have precise

location ability, and they are flexible for assigning

effective sensing and non-sensing fiber lengths

according to various application requirements. And

mature multiplexing and interrogation technologies

make the FBG-based sensing network a potentially

cost effective and promising monitoring system,

especially for perimeter of short or middle range.

Moreover, the FBG sensor response changes linearly

with the intrusion behavior impacts, so it is more

helpful to identify the event features. However, the

event identification is still a pending and challenging

problem due to the environmental complexity in

practical uses [20–22], such as changing climates

like wind and rain, and unpredictable wildlife

interferences. Even the same person could

introduce different effects due to different fence

materials or different sensor mount ways, which

presents the most difficult problem to extract

essentially distinguishable characteristics of the

event targets.

Thus in this paper, a smart perimeter intrusion

monitoring system based on the FBG strain sensor

network is addressed, in which it can not only

overcome the difficulties of detecting weak

intrusions from large amounts of nonequivalent

sensor nodes with high probability of detection (PD)

and low false alarm rate (FAR), but also it can

distinguish different threatening activities by using

the principal component analysis (PCA) feature

extraction in the time domain and the K-nearest

neighbor classifier.

2. Hierarchical detection and identification flow in a fiber-optical PIDS based on the FBG sensor network

A quasi-distributed fiber-optic PIDS based on the FBG sensor network is constructed as shown in Fig. 1. In this system, a huge number of FBG sensors with a certain central wavelength for each are connected in series or in parallel in the cable to act as the basic sensing segments. The cable can be attached to a physical fence or buried under ground to measure the mechanical deformation of the fence or other disturbances from outside. In the FBG interrogator, a light source with wide frequency band provides original optical signals, and simutaneously demodulates the optical signals reflected from the separate FBG sensors along the fiber and converts them into digital electronic signals. And a processing unit which acts as an alarm system, processes the signal array and decides if any threat happens along the protected perimeter, where it is and even what it is. The basic detection principle of the system is to monitor the shift of the returned “Bragg” wavelength due to the perturbations of the gratings. And the PD, FAR, and identification rate (IR) are the most concerned metrics for the system.

Fig. 1 Configuration of a fiber-optical PIDS based on the

FBG sensor network.

Generally, attacks on purpose only occur occasionally in the all day long monitoring, and for

most of the monitoring time there is actually no attack or intrusion. To detect the perturbation effectively with high confidence, a hierarchical

detection and identification method is introduced in this paper as shown in Fig. 2, which contains three

Huijuan WU et al.: Intelligent Detection and Identification in Fiber-Optical Perimeter Intrusion Monitoring System Based on the FBG Sensor Network

367

main stages: abnormity detection with high PD, nuisance alarm exclusion, and real threat identification and classification. At the first two stages we only focus on the detection aim and leave

the identification alone. Thus computation of the identification process can be neglected when there isn’t any threat, which is quite suitable for the

on-line monitoring.

Fig. 2 Hierarchical detection and identification flow for the FBG-based PIDS.

3. Abnormity detection with autocorrelation analysis and nuisance alarm exclusion

In the sensor network as shown in Fig. 1, all the

FBG sensors actually have different sensing

performances, due to their inherent different

sensitivities with slightly different productions and

packaging conditions, different attachings or

burying ways, and changing environments. The

sensing nodes in the network are nonequivalents and

the magnitudes of the acquired signal responses

differ greatly. Therefore, the traditional energy

thresholding method will definitely result in a low

PD or a very high FAR in practical applications.

And for a perimeter with fences of mixed materials,

the case will become even worse, and it cannot play

its role any more. Thus the authors introduced a new

solution, based on the different autocorrelation

characteristics between the attack signals and the

non-attack signals [19]. As shown in Fig. 3, the

non-intrusion and several typical intrusion signals

for the FBG strain sensors have distinguished

autocorrelation curves. The line with little square

marks represents the autocorrelation function of a

certain intrusion signal, such as climbing, striking,

swinging, and cutting signals, whose correlation lags

are always much longer than those of the lines

without square marks, which signify the case of the

regular signals without any external perturbation.

Here the time lag unit is sample with a sampling rate

of 500 Hz. From the autocorrelation curves, it can be

seen that the intrusion signal generated from a

determined energy source is always highly

correlated with itself, while the signals without

perturbation are always weakly correlated. Thus it

can be taken as a basic intrusion detection criterion

in the first step.

And in this paper, this method or criterion is also proved to be suitable for the FBG vibration sensor signals, which can be seen from Fig. 4. In Fig. 4, four

typical vibration signals of climbing, temperature rising, animal disturbing, and cutting are examined. The high autocorrelation characteristics of the

vibration signals are similar to those of the strain signals but they have more oscillation components for the same event signal.

At the detection stage, to detect the very weak signals from the sensor array, a lower correlation threshold is adopted to maintain a quite high PD.

Thus a lot of environmental interferences such as wind, rain, snow, and hailstone, could also be involved, which would be the main nuisance alarm

sources. It is thus necessary to exclude these nuisance alarms at the second stage. Fortunately, the environmental changes will influence almost all of

the sensors in the network, while real threats always interfere at a local area. And the local and global effects can be directly discriminated from the alarm

Photonic Sensors

368

sensor number and their locations. By excluding the frequently occurring nuisance sources, the event types to be discerned are significantly decreased,

which not only eases the computation load of the identification, but also makes the following identification much more reliable.

(a) (b)

(c) (d)

Fig. 3 Correlation characteristics of some typical FBG strain sensing signals: (a) climbing, (b) striking, (c) swinging, and (d) cutting.

4. Abnormity detection with autocorrelation analysis and nuisance alarm exclusion

4.1 Statistical feature extraction based on PCA analysis

As we investigate this problem, the structure or profile of the signals in the time domain reveals more distinguishable information for different types of events, which could be helpful for identifying certain threatening activities in the FBG-based opitcal-fiber fence. But the original temporal signal has a problem of data redundancy, which is

definitely not a good vector to be used as the input feature for the identification even though it contains the whole structure information. Here we use the PCA method to convert high dimensional data into a few principal features with much lower dimensions. It uses the temporal data which can tell the signal’s profile differences while avoiding its redundancy in the time domain, thus it could be a promising solution for the above problem. As a typical statistical data analysis method, the PCA is successfully employed to find the implicit modality buried in the redundant signals in face recognition [23] and fault diagnosis [24].


369

(a) (b)

(c) (d)

Fig. 4 Correlation characteristics of some typical FBG vibration sensing signals: (a) climbing, (b) temperature rising, (c) animal’s disturbing, and (d) cutting.

The principal components analysis process may

be regarded as a process of characteristic selection

following the feature extraction. First raw temporal

intrusion data are transformed into a feature space

with a much lower dimension, and its primary

statistical feature vectors are selected as its

characteristic bases. Each kind of training signal can

be approximated or reconstructed by a linear

combination of the primary feature bases. The

principal features are then selected and used to

profile the inputs without any redundancy. Assume

m different intrusion events happened, including

several typical activity patterns, such as climbing,

knocking, digging, and walking. The m detected

signals builds up a set of intrusion vectors, regarded

as the m dimensions random vector

1 2[ , , , ]TmX x x x . And its covariance XC which

is denoted as the covariance matrix in this paper is

defined as [( ( ))( ( ) )]T

XC E X E X X E X (1)

where E(X) is the expected mean value of X.

Calculate the eigenvalues 1 2, , , m and corresponding normalized eigenvectors

1 2, , , mU U U of XC : ( 1,2, , )X i i iC U U i m . (2)

If we consider the covariance XC as a vector of

m dimensions, and the eigenvectors 1 2, , , mU U U

decide the direction of the vector. The eigenvalues

Photonic Sensors

370

1 2, , , m are the contribution factors of each

eigenvector, respectively. If the eigenvalue

corresponding to the eigenvector is bigger, its

contribution radio is larger in reconstruction, and

vice versa. A few bigger eigenvalues are kept and

the smaller ones are neglected, thus the dimension

gets to fall. The fallen dimensions process selects

the principal components but discards secondary

ones, which is just the key spirit of the PCA method.

Supposing 1 2 m , the standard deviation

contribution radio of primary components is always

defined as

1 1

( )M m

i ii i

R M

. (3)

When the standard deviation contribution radio

( )R M is big enough (usually take above 90%), the

first M eigenvectors iU ( 1, 2, , )i M can be

selected to build up a projection space, which is also

regarded as the feature space. In this new space, the

inputs can be projected as

Ti iF X U ( 1, 2, , )i M . (4)

The primary features for each intrusion signal

are then selected and extracted, which can be

represented as a feature vector of M dimensions in

the fallen dimension space.

4.2 Identification using a K-Nearest neighbor classifier

Through the above transformation, the test

samples are individually projected in an

M-multi-dimensional feature space constructed by

the training samples. Because similar signals with

similar feature values could be located at a closer

location in the new space, then a K-nearest neighbor

classifier is used to discriminate them. However, the

identification rate (IR) of this method is mainly

dependent on whether the extracted PCA features as

above are distinguishable. It assigns a test sample to

the jth class if a majority of its nearest neighbors

belong to the jth class. The neighborhood is defined

using the Euclidean distance, in which the distance

between a test sample, xtest, and any training sample

x is given by

2

1

dist( , ) ( )M

i ii

xtest x xtest x

. (5)

The Euclidean distances describes the dissimilarity of the test sample and the train samples,

thus is always chosen as the dissimilarity representation (DR).

5. Experimental results and discussion

5.1 Hierarchical detection with autocorrelation analysis

In this section, the hierarchical detection with

the autocorrelation analysis is first investigated for

the fence-mounted sensing nodes. The experimental

setup is constructed as in Fig. 1. Thirty FBG strain

sensors are used, and the sensors are installed every

2 meters for monitoring a perimeter of about 65

meters length. The interrogator for demodulating the

FBG sensing signals is MOI si130 (MICRON

OPTICS, USA) with a sampling frequency of 500Hz.

In the test, five typical signals are included to be

detected, such as climbing, striking, swinging,

pushing, and non-threat signals. The test fences

include two kinds of wire mesh with different iron

materials and a kind of window bar of aluminum

alloy, which are remarked as Fence Type I, II, and

III, respectively. The test can be classified into two

groups: (1) The first group is for a single point

intrusion test, where a total of 564 weak personal

intrusions are exerted onto the three kinds of fences,

of which 178 are tested on the harder Al alloy bars

while the others are tested on the other two soft wire

meshes, with 198 intrusions each; (2) The second

one is for multiple-event detection where 350 tests

are run, and two events are simultaneously exerted

on two different fences at each time, thus there are

700 events in total for this group. The detection

results are concluded in Table 1. The PD can be

improved up to 99.65% and 99.57% for the single-

and multiple-event detection, respectively, and the

missing report rate mainly lies in the harder

aluminum alloy fence test group.


371

Table 1 Test results for single- and multiple-event detection based on the autocorrelation analysis.

Total test number Total events

Detected events

PD

Single event

564 (I: 150; II: 150;

III: 264) 564 562 99.65%

Dual events

350 (I: 100; II: 100;

III: 150) 700 697 99.57%

5.2 Identification with PCA analysis

5.2.1 Simulation results and discussion

To evaluate the effectiveness of the proposed

feature extraction and identification method,

different simulated signals are first trained and

tested. As shown in Fig. 5(a), five different types of

signals are generated and taken as the training

samples, which include two sinusoidal signals of

10 Hz and 1 Hz, respectively, two periodic square

waves with different periods (Period1=10 s,

Period2=5 s, both with duty cycle of 50%) and a

periodic triangular wave with a period of 6 seconds.

For the testing samples, we add different delays to

each kind of signal and modify the amplitude for

each delayed version as shown in Fig. 5(b), for being

close to real cases. A total of 30 testing samples are

generated corresponding to the five training samples

above, with 6 testing samples for each type. As we

can see from the identification results in Fig. 6, an IR

is acquired as high as 96.67% in the simulation test,

which proves that different signals’ profile or

structure can be extracted by PCA, and it could

make a good classification result.

5.2.2 Field test results and discussion

To test its actual effectiveness, the field test for

the FBG-based fiber optical PIDS is also carried out

around a school building, in which both the

fence-mounted and the ground-buried sensors are

tested. The sensor array and the signal demodulating

configuration used is the same as above in the

detection test in Section 5.1. Most of the sensors are

attached onto the perimeter fence, and only two of

them are buried under the ground. For the

ground-buried type, to enhance its sensitivity in a

(a)

(b)

Fig. 5 Training and testing sample signals for PCA analysis and identification: (a) five kinds of training sample signals and (b) testing signals corresponding to the training samples.

Fig. 6 Identification results for the simulation test (IR=96.67%).

larger area, we bond the sensing cable onto a soft

wire mesh of 80 cm width, and bury them into the

ground together with a burial depth of 15 cm in the

clay soils. Four seconds of data are taken as a

processing unit and analyzed for the event decision,

Photonic Sensors

372

which is updated at each second. A field database is

constructed for eight typical events as shown in

Fig. 7, which includes 1322 intrusion signals in total.

Four types are for the fence-mounted sensing cable,

Fig. 7 Typical event signals in the field test.


373

e.g. personal climbing on the fence, personal

striking on the fence, heating on the fence, cat

jumping on the fence, and the other four are for the

ground-buried cable, such as personal digging on the

ground, personal walking on the ground (across the

cable), personal running on the ground (across the

cable), and cat crawling on the ground. Half of the

signals in each type are taken as training samples

(661 signals in total), and the other half are testing

samples (661 signals in total).

Figure 7 shows that the differences of the eight

event signals mainly lie in their varying profiles of

time sequences. Their averaged PCA feature values

are extracted and concluded in Table 2. From the

first four dimensions of the feature space, it can be

seen that each kind of event has distinguishable PCA

feature values especially in the 1st dimension.

Choosing the feature dimension as M =4, most of

the identification results can be achieved up to a

hundred percent as shown in Table 3, except two

events occurring in the ground-buried fences,

walking and running, due to the essential similarity

between these two activities. In general, the average

IR can be achieved up to 96.52% for the eight

typical event targets.

Table 2 PCA feature values for eight typical event signals in the field test.

1st

dimension 2nd

dimension 3rd

dimension 4th

dimensionPersonal climbing

~68575.61 ~ –27.83 ~0.39 ~ 0.44

Striking ~68710.36 ~ –26.97 ~ –0.12 ~ 0.27 Fire ~ 67576.74 ~ –26.69 ~ 0.02 ~ 0.26

Cat jumping on the fence

~ 68562.35 ~ –27.05 ~ 0.22 ~ 0.23

Digging ~ 69000.15 ~ –27.32 ~ 0.12 ~ 0.22 Walking ~ 69261.58 ~ –27.29 ~ 0.08 ~ 0.32 Running ~ 69128.55 ~ –27.34 ~ 0.04 ~ 0.27

Cat crawling on the ground

~ 68991.10 ~ –27.04 ~ 0.42 ~ 0.67

The average IR and its elapsed time varying with

the dimension of the PCA features are also

investigated in Fig. 8 for the proposed feature

extraction and event identification method in this

field test. As we can see that the IR is always kept

above 90%, even though it slightly changes with the

PCA feature dimension M and performs best when

M is equal to 4, which gives a good proof for the

proposed method based on the PCA feature

extraction. The elapsed time for the 661 test samples

takes more or less 1.2 seconds, and each processing

unit’s computation time takes less than

2 miliseconds. The processing time can be nearly

negligible thus the algorithm is very suitable for the

on-line identification and classification. Table 3 Identification results for eight typical events in the

field test.

IR(%) when M=4 is

chosen Personal Climbing

(74 events) 100%

Knocking (25 events)

100%

Firing (109 events)

100% Fence-mounted

Cat jumping on fence

(23 events) 100%

Digging (121 events)

100%

Walking (98 events)

89.8%

Running (191 events)

93.2% Ground-buried

Cat crawling on the ground

(20 events) 100%

Average IR of

96.52%

Fig. 8 Average IR and its elapsed time varying with the

feature dimension M.

6. Conclusions

In this paper, it is first presented that the PCA

method can be used to extract different intrusion

signal’s profiles or structures in the time domain and

give good features for the following behavior

identification of the FBG-based fiber-optical PIDS.

Based on the autocorrelation characteristics analysis

and by using a hierarchical detection and

identification model, the PD for multiple-event

detection can even be improved up to 99.57%, and

Photonic Sensors

374

an average recognition rate for eight typical events

in real field test can be achieved as high as 96.52%,

which is suitable for both the fence-mounted and the

ground-buried applications. Moreover, the proposed

detection and identification method can be carried

out on line in real time. The good performance or

intelligence improvement of the proposed method

can promote its application in many important areas

in perimeter security, safety monitoring of oil/gas

pipe lines, electrical power lines, large-scale civil

structures, etc.

Acknowledgement

The authors gratefully acknowledge the previous

supports provided by National High Technology

Research and Development Program of China (863

Program, Grant No. 2007AA01Z245), the supports

provided for this research by the Major Program

(Grant No. 61290312) and Youth Foundation (Grant

No. 61301275) of the National Natural Science

Foundation of China (NSFC), and the Fundamental

Research Funds for the Central Universities (Grant

No. ZYGX2011J010). This work is also supported

by Program for Changjiang Scholars and Innovative

Research Team in University (PCSIRT, IRT1218),

and the 111 Project (B14039).

Open Access This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.

References

[1] J. Geng, Y. Zou, S. Staines, M. Blake, and S. Jiang, “A real-time distributed fiber strain sensor for long-distance perimeter intruder detection,” in Optical Solutions for Homeland and National Security (Optical Society of America), U. S. A., pp. P3, 2005.

[2] M. Szustakowski, W. Ciurapinski, N. Palka, and M. Zyczkowski, “Recent development of fibre optic sensors for perimeter security,” in Proceedings of the International Conference: Modern Problems of Radio Engineering, Telecommunications and Computer Science, Chile, pp. 158–162, 2002.

[3] S. V. Shatalin, V. N. Treschikov, and A. J. Rogers, “Interferometric optical time-domain reflectometry for distributed optical-fiber sensing,” Applied Optics, 1998, 37(24): 5600–5604.

[4] J. Bush, C. A. Davis, P. G. Davis, A. Cekorich, and F. P. McNair, “Buried fiber intrusion detection sensor with minimal false alarm rates,” in Proc. SPIE, vol. 3489, pp. 285–295, 1998.

[5] J. C. Juarez, E. W. Maier, K. N. Choi, and H. F. Taylor, “Distributed fiber-optic Intrusion sensor system,” Journal of Lightwave Techology, 2005, 23(6): 2081–2087.

[6] J. C. Juarez and H. F. Taylor, “Field test of a distributed fiber-optic intrusion sensor system for long perimeters,” Applied Optics, 2007, 46(11): 1968–1971.

[7] Y. J. Rao, J. Luo, Z. Ran, J. Yue, X. Luo, and Z. Zhou, “Long-distance fiber-optic Φ-OTDR intrusion sensing system,” in Proc. SPIE, vol. 7503, pp. 75031O-1–75031O-4, 2009.

[8] F. Peng, Z. Wang, Y. Rao, and X. Jia, “106 km fully-distributed fiber-optic fence based on P-OTDR with 2nd-order Raman amplification,” in Optical Fiber Communication Conference and Exposition and the National Fiber Optic Engineers Conference (OFC/NFOEC), Anaheim, CA, pp. 1–3, 2013.

[9] A. D. Kersey, “Recent progress in Interferometric fibre sensor technology,” in Proc. SPIE, vol. 1367, pp. 2–12, 1990.

[10] A. A. Chtcherbakov and P. L. Swart, “Polarization effects in the Sagnac-Michelson distributed disturbance location sensor,” Journal of Lightwave Technology, 1998, 16(6): 1404–1412.

[11] B. Kizlik, “Fibre optic distributed sensor in Mach-Zehnder interferometer configuration,” in Proceedings of the International Conference: Modern Problems of Radio Engineering, Telecommunications and Computer Science, Chile, pp. 128–130, 2002.

[12] M. Szustakowski, W. M. Ciurapinski, and M. Zyczkowski, “Trends in optoelectronic perimeter security sensors,” in Proc. SPIE, vol. 6736, pp. 67360Q-1–67360Q- 12, 2007.

[13] L. Yuan and Y. Dong, “Loop topology based white light interferometric fiber optic sensor network for application of perimeter security,” Photonic Sensors, 2011, 1(3): 260–267.

[14] A. D. Kersey, M. A. Davis, H. J. Partrick, M. Leblance, K. P. Koo, C. G. Askins, et al., “Fiber grating sensors,” Journal of Lightwave Technology, 1997, 15(8): 1442–1463.

[15] Y. Rao, “In-fibre Bragg grating sensors,” Measurement Science and Technology, 1997, 8(4): 355–375.

[16] Y. Rao, “Recent progress in application of in-fiber


375

Bragg grating sensors,” Optics and Lasers in Engineering, 1999, 31(4): 297–324.

[17] Z. Ran, Y. Rao, N. Nie, and R. Chen, “Long-distance fiber Bragg grating sensor system based on hybrid Raman/erbium-doped fiber amplifier,” in Proc. SPIE, vol. 5855, pp. 583–586, 2005.

[18] Q. Jiang, Y. Rao, and D. Zeng, “A fiber-optical intrusion alarm system based on quasi-distributed fiber Bragg grating sensors,” in APOS’08. 1st Asia-Pacific Optical Fiber Sensors Conference, Chengdu, China, pp. 1–4, 2008.

[19] H. Wu, Y. Rao, C. Tang, Y. Wu, and Y. Gong, “A novel FBG-based security fence enabling to detect extremely weak intrusion signals from nonequivalent sensor nodes,” Sensors and Actuators A: Physical, 2011, 167(2): 548–555.

[20] F. Blackmon and J. Pollock, “Blue Rose perimeter defense and security system,” in Proc. SPIE, vol.

6201, pp. 620123, 2006. [21] D. Anderson, “Smart perimeter security,”

Fiber-SenSys.,http://www.fibersensys.com/index.php?option=com_docman&task=doc_details&gid=55&Itemid=54 (2009).

[22] H. Yan, G. Shi, Q. Wang, and S. Hao, “Identification of damaging activities for perimeter security,” in 2009 International Conference on Signal Processing Systems, Singapore, pp. 162–166, 2009.

[23] Y. Zhao, X. Shen, N. Georganas, and E. M. Petriu, “Part-based PCA for facial feature extraction and classification,” in IEEE International Workshop on Haptic Audio visual Environments and Games, Lecco, pp. 99–104, 2009.

[24] X. Han, D. Xu, and Y. Liu, “Application of principal components analysis in condenser fault diagnosis,” in The Sixth World Congress on Intelligent Control and Automation, Dalian, pp. 5666–5669, 2006.

Documents

Intelligent Detection and Identification in Fiber-Optical ... · Huijuan WU et al.: Intelligent Detection and Identification in Fiber-Optical Perimeter Intrusion Monitoring System